SIMA 2 is an agent that can play, reason, and learn together with users in virtual 3D worlds. It builds on 15 years of AI research in games, integrating reinforcement learning, hierarchical planning, and natural‑language interaction.
The November 2025 research paper demonstrates SIMA 2’s transfer‑learning capabilities across platforms ranging from classic Atari titles to the massive multiplayer online game EVE Online. The system follows a perception‑decision‑execution pipeline: a convolutional neural network extracts visual features, a Monte Carlo Tree Search‑based hierarchical planner generates strategies, and an action‑mapping module translates high‑level commands into low‑level controls.
Results show that when collaborating with human players, SIMA 2 significantly improves task success rates and exhibits expert‑level reasoning in complex strategy games, laying a solid foundation for future general‑purpose game AI.
Review